Some notes on your robot's odometry:
* The yaw estimate is pretty bad, exhibiting a very high rate of drift.
* It feels like there's a non-zero bias in the yaw integration
(either under- or over-reporting of rotation), but I was unable to
compensate for it, so I'm not sure of this.
* Jumps in the robot's pose suggest instances of non-trivial wheel
slip; do you see the wheels slip against the floor?
To get the behavior shown in the video, I adjusted three parameters in amcl (with respect to amcl_diff.launch):
* increase odom_alpha1 to 0.8, which is used to compute uncertainty
in rotation and translation, as a function of rotation.
* increase odom_alpha2 to 0.8, which is used to compute uncertainty
in rotation, as a function of rotation
* decrease update_min_a to 0.2, which is the change in angle
required to cause a filter update